eyegazer-vit-binary / README.md
monsoon-nlp's picture
Update README.md
25f706b
metadata
library_name: peft
base_model: google/vit-large-patch16-224

LoRA Image Binary Classification LoRA adapter

Trained on APTOS 2019 Kaggle competition for identifying diabetic retinopathy. In this case I've modified the problem to binary classifier (diagnosis=0 vs. all others; 50-50% distribution in training data)

Base Model: google/vit-large-patch16-224

Dataset: https://www.kaggle.com/c/aptos2019-blindness-detection - fundus images of the back of the eye, and diabetic retinopathy score

Training notebook: https://colab.research.google.com/drive/1TVsUyyou87E26Sz40CdBH3CzWoVckgtq?usp=sharing

On 10% held-out of training data: accuracy 98%

  • PEFT 0.5.0

PEFT Image classifier inference / Gradio app

from peft import PeftModel
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForImageClassification

from torchvision.transforms import (
    CenterCrop,
    Compose,
    Normalize,
    RandomHorizontalFlip,
    RandomResizedCrop,
    Resize,
    ToTensor,
)

model_name = 'google/vit-large-patch16-224'
adapter = 'monsoon-nlp/eyegazer-vit-binary'

image_processor = AutoImageProcessor.from_pretrained(model_name)

normalize = Normalize(mean=image_processor.image_mean, std=image_processor.image_std)
train_transforms = Compose(
    [
        RandomResizedCrop(image_processor.size["height"]),
        RandomHorizontalFlip(),
        ToTensor(),
        normalize,
    ]
)

val_transforms = Compose(
    [
        Resize(image_processor.size["height"]),
        CenterCrop(image_processor.size["height"]),
        ToTensor(),
        normalize,
    ]
)

model = AutoModelForImageClassification.from_pretrained(
    model_name,
    ignore_mismatched_sizes=True,
    num_labels=2,
)

lora_model = PeftModel.from_pretrained(model, adapter)

img = Image.open("sample.png")
pimg = val_transforms(img.convert("RGB"))
batch = pimg.unsqueeze(0)
op = lora_model(batch)
vals = op.logits.tolist()[0]

if vals[0] > vals[1]:
    return "Predicted unaffected"
else:
    return "Predicted affected to some degree"

Future goals

  • More documentation
  • Modify loss for regression on 0-4 score